SMA

An opinion holder/source and time of opinion.

GENERAL

An opinion summary is just like a multi-document summary because we need to summarize multiple opinion documents, e.g., reviews. • It is, however, very different from traditional multi-document summary.

The core form of opinion summary is called the aspect-based opinion summary.

In an application, the number counts is often replaced by percentages (quantitative perspective is especially important in practice).

Polarity classification: is something negativ, neutral or positive,

Valence and arousal: The Valence-Arousal-Dominance (VAD) model is a psychological framework that aims to represent and measure the three fundamental dimensions of emotional experience: valence (+/-), arousal(intensity), and dominance.

• Dominance: Indicates the level of control or influence associated with an emotion
• It ranges from low dominance (submissive) to high dominance (dominant)

Subjectivity and objectivity: The task of allocating a given text (usually a sentence) into one of two categories: subjective or objective. Sometimes, it can be more difficult than polarity classification problem. For example, the subjectivity of words or phrases usually depends on the context, but an objective chapter such as a news article may quote people’s opinions which are subjective sentences. It could be helpful to remove objective sentences from a document before classifying its sentiment polarity.

Beyond polarity: fine-grained opinion analysis aims to identify types of opinion entities such as opinion holders, opinion expressions, opinion targets, aspects of a target, opinion sources.

Feature or Aspect-based: Identifying relevant entities. Extracting their features or aspects. Determining whether an opinion expressed on each aspect or feature is positive, negative, or neutral.

Syntactic ambiguity: sentence or phrase can be parsed in multiple ways due to the structure of the language, leading to different possible interpretations.

• Example: They ate pizza with anchoves (did they eat pizza with anchoves on top or did they eat pizza using anchoves?)

Semantic ambiguity:  occurs within a language (e.g., the word “organ” in English means both a body part and a musical instrument), but it can also cross a language boundary, such that a given word form is shared in two languages, but its meanings are different (e.g., the word “angel” means “sting” in Dutch). --> In linguistica, un'espressione è semanticamente ambigua quando può avere molteplici significati. Maggiore è il numero di sinonimi di una parola, maggiore è il grado di ambiguità.

Pragmatic ambiguity: it arises from the nuances of meaning, implications or specific contexts of a communication

• Example: Do you know what time it is? One actually wants to know what the time is.One expresses anger about someone being late for an  appointment.

Emotional ambiguity: The same expression can be interpreted with its signification or with its opposite meaning than, for example, the irony used.

• Example: Today is Marco Viviani's lecture on Sentiment Analysis, can't wait (can't wait could be meant iterally, or ironically)

Knowledge-based techniques:

• We create the lexicon
• popular because it is accessible and affordable
• classifies text into affect categories like happy, sad, etc.
• poor recognition of affect when linguistic rules are involved --> "today wasn't a happy day at all" might be categorized as happy 

Supervised methods:

• Machine and deep learning
• We give some labels and then have the algorithm learn to label the rest itself

Hybrid approaches:

• Mix of knowledge-bases and supervised methods
• Given lexicons can be used as additional features in supervised approaches to improve the effectiveness of the considered model.

General Inquirer: It has sentiment labels for about 3,600 terms. 

AFINN:

• Scores words on a scale from -5 to +5. A negative score indicates a negative sentiment, a positive score indicates a positive sentiment. 
• Limited to unigrams (single words)

SentiWordNet: SentiWordNet is a sentiment lexicon which augments WordNet with sentiment information.

SenticNet: Has sentiment entries for 30,000 words and multi-word expressions using information propagation to connect various parts of common-sense knowledge representations.

• Valence Aware Dictionary and sEntiment Reasoner is a pre-built sentiment analysis tool designed to analyze the sentiment of text data, such as sentences or paragraphs. 
• Lexicon based 
• sentiment polarity and intensity
• good for social media: understands slang, Big or small letters, certain punctuation (in this case we should not delete emojis, punctuation here, otherwise it does not get the fulll picture)

Examples:

LIWC: analyzes written or spoken language based on linguistic and psychological dimensions.

WordNet-Affect: annotates senses in WordNet with emotions. It consists of 2,874 synsets annotated with affective labels (called a-labels).

EmoLex: words are categorized into different emotional categories, such as joy, anger, sadness, fear, disgust, and surprise.

SenticNet: Multilingual support, publicly available

SemEval

Standford NLP Group

Amazon review dataset

Yelp review dataset

Observable communities (explicitly visible):

• Explicit user groups: when people join groups or pages focus on specifi topics, etc.
• Hashtags and topics: communities defined by using common hashtags or engagement with specific topic
• Friendships and connection

Latent communities (inferred through data analysis):

• Network analysis 
• Clustering algorithms
• Topic modelling

• Discover functionally-related objects: Detect communities (e.g. group related to gaming)
• Study interaction between groups: e.g. political field --> Two groups (one left leaning, one right leaning) and see what are their connection points
• Infer missing nodes values: There is a node (of which we don't have the age for example) --> we look at which community it belongs to and can estimate the age based on the average age of the community
• Predict unobserved connections: Also called Link prediction (for example, linked in --> they understand to which community we belong to and based on that make us suggestion, of who we could know "Vielleicht kennst du…")

Some reasons for popularity-based preferential attachment include:

• Visibility: popular nodes are more visible and easily discoverable, making them more likely to receive new connections.
• Social proof: people tend to trust and follow nodes that already have many connections or followers, assuming that they provide valuable information or resources.
• Network effects: in some networks, the value of being connected to a popular node increases as more users connect to it, reinforcing its popularity.

Reasons for quality-based preferential attachment include:

• Content excellence: nodes that consistently produce high-quality content or services attract more connections because users appreciate the value they provide.
• Trustworthiness: quality nodes are trusted sources of information or services, which encourages more users to connect to them.
• Longevity: nodes that maintain high quality over time tend to build a loyal user base, leading to further connections.

Hubs are expected to link to each other (by the virtue of the many links they have) --> celebrity couples

However, in some networks they do, in others they do not, and hubs avoid hubs! (e.g. protein networks)

Assortative networks have been observed mostly in social networks, whereas many technological and biological networks display to be disassortative.

Mathematically, assortativity is defined as the Pearson Correlation Coefficient. 

Positive values indicate a correlation between nodes of similar degree, while negative values indicate relationships between nodes of different degree.

When r = 1, the network is said to have perfect assortative mixing patterns.

When r = 0, the network is non-assortative (or neutral).

When r = −1, the network is completely disassortative.

• A connection metric
• Tendency of individuals to associate and tie with other similar ones
• Opposite of heterophily
• social media favours homophily by suggesting to a user similar post to the one he already liked or interacted with

Filtering algorithms: algorithms that study behaviour of user in the past. If he liked and joyed certain topics/posts, the algorithm will tend to provide the user with similar posts/products.

Personalization: Profile is a list of interest that a user has. If the user makes searched on SM plattform than the user receives only information that are similar to his profile. 

Positive:

• interpersonal similarity improves coordination and increases the expected payoff of interactions, beyond the mere appreciation of others.
• helps people access information, innovations and widespread behaviors, and forms opinions and social norms.
• Individuals are more likely to successfully influence others when they are similar to them.

Negative:

• Can translate into limited social worlds, with strong implications for how information is received and disseminated, and the attitude and form of interactions people experience.
• Causes filter bubbles (algorithm shows mostly things you like and agree with) which leads to echo chambers (people see only content that they agree with and interact only with other people in these chambers)